I enjoyed the letter from Dr. Craig Reeves (June 17th). He argues that higher education institutions are consciously choosing not to deal with widespread fraud using generated AI to avoid sacrificing income from international students. He is right that international students are supporting universities in the UK. However, it is not true that universities can simply spot AI fraud if they wish. Dr. Reeves says that AI detectors should be used, but the research he cites rebuts this argument.
The last study he cited (Perkins et al, 2024) shows that AI detectors were accurate in less than 40% of cases, which dropped to just 22% of “hostile” cases when AI use was intentionally obscured. In other words, the AI detector could not find that AI was in use in three quarters at the time.
So it's wrong to say there is a simple solution to generative AI problems. Some universities have courageously pursue academic misconduct cases against students using AI. However, since AI does not leave traces, it is almost impossible to clearly indicate that a student has used AI unless the student acknowledges it.
In the meantime, the institution has switched to “safe” ratings, including in-person exams he celebrates. Others design ratings assuming that students use AI. No one says that the university is right. However, if confusion is a simpler explanation, then conspiracy should not be envisaged.
Josh Freeman
Policy Manager at the Institute for Higher Education Policy Research. author, Student Generated AI Survey 2025
Using AI to “write” things in higher education encourages important research and discussion at institutions, and accurate reporting of that research is clearly important. Craig Reeves mentions three papers in favour of the Turnitin AI Checker, claiming that the university opted out without testing this feature because of concerns over false positive flagging of AI-generated human-generated texts. One of these papers states: “Researchers conclude that the available detection tools are neither accurate nor reliable, and have a major bias for classifying the output as human-generated, rather than detecting text generated by AI,” (Weber-Wulff et al.). Also, the second Turnitin was the worst of the seven AI detectors tested to flag AI-generated text, with 84% not detected (Perkins et al). AI detectors can easily avoid false positives by not flagging text.
Clicking on almost infinite, superficially plausible texts requires careful thoughts about how you evaluate your work.
Professor Paul Johnson
University of Chester
Otherwise, in the obvious (and perhaps convenient) well-thought-out criticism of what blind spot higher education has for AI use, Craig Reeves appears to encourage a return to traditional exams as a way to eradicate the problem.
I sympathize (and strongly believe that something should be done), but I hope this doesn't happen in a way that “fits all sizes.” I have been marking the exams for over 30 years. During that period, I was regularly impressed by the students' understanding of topics. I remember that I only enjoy reading one exam essay. Others, no matter how good they are, read like a delusional flow of consciousness. The central transferable skills offered by a humanities degree is the ability to write well on a specific topic after research. The exam cannot provide it.
I seek a move towards more analytical assessments that students face new material that they must consider in a short period of time. I think leaving the traditional essay as the only form of evaluation might help reduce the effect of external input (of course, it doesn't stop). From experience, this focus also helps students move towards applying new understandings rather than passive digestion of ideas.
Professor Robert McColl Miller
Chairperson of Linguistics and Scottish Language at the University of Aberdeen
